AlphaGazette
Menu

Finance

Daisytuner raises €6 million seed for cross-chip AI inference

The Darmstadt firm also released Daisyflow, which customers can host themselves, and says it beat vLLM by up to 18 per cent on an AMD GPU test with Qwen 2.5.

1 min read 2 sources AI-generated

Daisytuner, based in Darmstadt, has closed a €6 million seed round and released Daisyflow, software that tunes models while they answer requests rather than while they train. Elaia and LEA Ventures led together. Seedcamp and the Amadeus APEX Technology Fund also put money in. Operators can place Daisyflow inside facilities they control, instead of depending on an external AI cloud.

The stack pairs a compiler that learns from its own results with a patented platform for optimisation. DOCC, the open-source compiler, turns PyTorch and other widely used framework code into native programs for GPUs from NVIDIA and AMD and for newer chip designs. Agents try candidate speed-ups continuously on physical hardware. Each result that checks out is kept and reused when similar code shows up later.

In a test the company ran, Daisyflow served Alibaba's Qwen 2.5 on AMD Instinct GPUs faster than vLLM, a widely used open-source inference engine, by up to 18 per cent. Daisytuner says the capital will go toward more business customers, further work on the compiler and additional hires.

Lukas Trümper and Adrian Schmitz founded the startup in 2024. The group draws on over a decade of research that includes Lawrence Livermore National Laboratory, ETH Zurich, TU Darmstadt, RWTH Aachen and TU Munich. Sébastien Lefebvre, a partner at Elaia, said one codebase should reach full speed on the silicon a buyer chooses, with no rewrite.